GARDAIS Ionel wrote:
Beside the fact that the hit rate is low, response time are way too long for users (cache-misses median service times are around 200ms and cache-hits are around 3ms)
---- Have you tried access without a squid-proxy -- but maybe using 'socks' (if you need to bridge a firewall), or direct? I've done performance analysis on my squid installation a few times -- cuz not happy with performance...but.. when squid isn't there and I go through socks, or I setup a machine on the net directly -- I don't get any faster throughput or response time ... at most, we're talking differences down in the noise level ... One-by-one, I eliminated or upgraded parts that I could -- my internal net is 1Gigabit. Tried upgrading to a firewall product that supported a 100Mb interface (old was 10Mbit&half duplex)...but its a matter of my DSL not being much faster than yours -- ~2.5-3Mb. But simple 'ping' times to most servers are 100ms or more. within my ISP's network, they are as low as 50-55, but going out, and back down someone else's network usually adds at least another 50, often nearer 100, so just ping times alone to many outside destinations are 100-150, closer to 150 on heavier sites. So if a tiny ping packet -- handled by the OS takes that long -- of course adding web-server application time to that is going to add 'something'. So while some outside servers can serve 1-segment packets in the 150-160ms range, most are higher ... A TCP_REFRESH_UNMODIFIED might take around 140-150... But most misses are 400, 600 or more, easily. My hits are fast...some coming in at less than 1ms (0 is stated), but if you have to go to a webserver -- forget it...300-1000 is typical -- *without* a cache (in my case, anyway). The only ways to speedup web pages are to make sure your clients are issuing multiple requests in parallel -- easily configured in Firefox, but there are KB articles (& pay-utils that will set the values for you, of course) for MS products. Since I take it that pipeline requests doesn't work in squid(?) and compression doesn't work (not that compression would help latency -- only throughput), the best thing is to make sure your clients can issue at least 8 requests in parallel at a time -- through your squid proxy and that your squid-proxy can handle that many connections/client without becoming a bottleneck. My proxy usage is often just me -- so no prob, but even when I have housemates and guests active, squid never seems to be a bottleneck.
